Internal model validation, use test and ALM assumptions
You are the chief risk officer of an Indian insurer. Your internal model, used in the ORSA, produces a 1-in-200 one-year VaR of ₹6.4bn. Catastrophe risk is ₹2.1bn of that, driven by a vendor model whose India flood vulnerability function was calibrated on 2015-2020 data. Validation found that 23% of exposure locations are geocoded only to district level and that the model use test is weak: the three-year plan assumes flat 10% growth and the model output is not used in reinsurance purchasing. The investment portfolio has a duration of 4.2 years against a liability duration of 6.8 years, and the model assumes a 6.5% yield throughout. A 1-in-20 stress produces a ₹1.1bn loss, taking the solvency ratio from 1.85 to 1.40. The regulator has asked for a submission under the proposed risk-based capital framework.
1. Critically evaluate the model's limitations and set out a validation plan before submission.
Read the worked answer
The model has four weaknesses. First, data quality: 23% of locations are geocoded only to district level, so the cat model cannot distinguish flood exposure within a district. That can understate or overstate concentration and tail risk. Second, calibration: the flood vulnerability function uses 2015-2020 data, a short period that may not capture recent urbanisation, drainage changes or extreme events. Third, the use test: the model is not used in reinsurance purchasing and the plan assumes flat growth, so the model is not informing decisions. Fourth, the ALM assumption: a constant 6.5% yield is not a scenario; it ignores reinvestment risk and the duration gap.
Validation plan: reconcile exposure data to policy and sum-insured records; improve geocoding for the largest locations; run sensitivity tests on vulnerability and on the 23% district-level exposure; backtest the cat model against the 2024 floods and other events; benchmark against vendor and industry views; test dependency assumptions between cat and reserve risk; and document expert judgement. Have the validation reviewed independently. For the use test, show how model output changes a real decision, for example the cat retention or the reinsurance purchase. If the model is not fit for use, submit a standard-formula calculation with a documented limitation rather than overstate confidence.
2. Advise the board on whether to submit the model, how to use it in the ORSA, and how to treat the 1-in-20 stress.
Read the worked answer
Submit only with a clear statement of limitations and a remediation plan. A partial internal model can be used if it meets tests of data quality, methodology, calibration, validation and use. The board should not treat the ₹6.4bn VaR as precise. Present a range and the sensitivity to the cat model, geocoding and the yield assumption. If the cat module is not validated, use a conservative add-on or fall back to the standard formula for that risk.
In the ORSA, link the model to decisions: the reinsurance programme, the investment strategy, the dividend policy and the growth plan. The use test is not a documentation exercise; show that the board changed a decision because of the model. For example, test whether the second cat layer is worth buying against the model's net loss curve.
The 1-in-20 stress takes the ratio from 1.85 to 1.40, below the 1.50 regulatory minimum. That is a serious capital adequacy signal even though it is a 1-in-20 event. The board should hold a buffer above 1.50 and identify management actions: reduce exposure, buy reinsurance, raise capital or de-risk assets. Do not average the stress away. Report it alongside the 1-in-200 and explain the difference between the regulatory minimum and the internal risk appetite.
3. Discuss the ALM gap and the 6.5% yield assumption, and explain how they affect the model and the board's decisions.
Read the worked answer
The portfolio duration is 4.2 years and the liability duration is 6.8 years, a gap of 2.6 years. If yields rise by 100bp, asset values fall by about 4.2% and liability values fall by about 6.8%, so the surplus improves on a duration basis. If yields fall by 100bp, the opposite happens and the surplus falls. The insurer is therefore exposed to falling interest rates, which is a real risk for long-tail liability reserves. The gap also creates reinvestment risk: assets mature before liabilities and must be reinvested at lower yields.
The model's constant 6.5% yield is not a stress or a scenario. It ignores the possibility that yields stay low, that credit spreads widen, or that the portfolio is illiquid. If the model discounts liabilities at 6.5% while the market yield is lower, it may understate liabilities and overstate capital. The board should see the surplus under at least three yield scenarios: unchanged, −100bp and +100bp, with matching asset and liability revaluations.
For decisions: the ALM gap supports buying longer-dated assets or using derivatives to extend duration, but that must be weighed against credit and liquidity risk. The board should also consider whether the liability duration is stable; if claims are paid faster or slower, the gap changes. The 1-in-20 stress and the ALM gap should be presented together, because a falling-yield scenario plus a reserve deterioration is a plausible combined stress.